When you train the machine learning model for face recognition, make sure you include sufficient data. Especially include dark-skinned images with multi-color and light backgrounds. However, make sure you also include fair-skinned images as well. It would also be a good approach to manually sort pictures according to pigments, then create multiple models for each set. This way, you can follow a multimodal approach, which I feel could improve results.

DR Craig Could you kindly re compose your question? Thanks.

Why do facial recognition systems struggle to provide accurate results, when matching people with darker pigmented skin?

It should be obvious: darker skin results in a lower contrast, making feature extraction more difficult. Thus you have either less or less accurate features.

And whether AI or classical math: "garbage in, garbage out".

Thank you Dreher. However I would argue that if it's that simple amplification could be used to resolve this issue inorder upscale systems and provide more accurate results. How could A.I be used to address this is issue?

It's most likely not "simple amplification" - more like "contrast enhancement" or "image normalization" prior further processing.

Although this issue could be addressed by classical image processing, one might develop some AI solution to prepare (normalize) the raw input in a way that subsequent facial recognition yields better results.

The dark side: no, I have no ready-made solution at hand. So it will be up to you to device some method for image normalization :)

Ok thanks that's, perfect. I'm actually doing a research paper so mostly theoretical but I'm working on a program in python using Rekognition as a platform. I'm hoping I will be able to normalize images to improve results from the scans on darker pigmented skin. Do you think this could work?

When you train the machine learning model for face recognition, make sure you include sufficient data. Especially include dark-skinned images with multi-color and light backgrounds. However, make sure you also include fair-skinned images as well. It would also be a good approach to manually sort pictures according to pigments, then create multiple models for each set. This way, you can follow a multimodal approach, which I feel could improve results.